探索用大型语言模型创建的合成医疗数据集的检测方法
Andrea Taloni1,2,3,4, Giulia Coco5, Marco Pellegrini1,2,3
1Department of Translational Medicine, University of Ferrara, Ferrara, Italy.
JAMA ophthalmology
|April 24, 2025
概括
像GPT-4这样的复杂的人工智能模型可以创建可能看起来真实的合成医疗数据集,对科学完整性构成风险. 需要进一步的研究来检测和防止AI驱动的数据制造在研究中.
科学领域:
- 人工智能的人工智能
- 医疗信息学 医疗信息学
- 数据科学数据科学数据科学
背景情况:
- 大型语言模型 (LLM),如生成预训练变压器4 (GPT-4),已经证明了产生合成医疗数据集的能力.
- 这些合成数据集可以设计为支持伪造的科学证据,引发对研究完整性的担忧.
研究的目的:
- 调查统计模式,表明数据制造的LLMs.
- 探索改进合成数据集的方法,以逃避真实性检查并提高其可靠性.
主要方法:
- 针对使用GPT-4o和自定义GPT模型的三个虚构临床研究,生成了合成数据集.
- 对未精炼和精炼数据集进行了法医分析,以确定统计异常和不切实际的临床记录.
主要成果:
- 最初的法医分析显示,在未经精炼的数据集中,有许多制造标志 (33.9%),包括性别名称不匹配,周末访问日期和年龄计算错误.
- 精制的数据集显示,制造迹象显著减少 (4.6%),其中四个数据集通过了法医分析.
- 然而,一些精细的数据集仍然表现出可疑的特征,例如不寻常的分布形状.
结论:
- 先进的定制GPT模型可以生成可能通过法医审查的合成数据.
- 人工智能制造出看似真实的数据集的潜力需要强大的检测方法来保持科学有效性.
相关概念视频
02:55Synthetic Biology
4.7K
Synthetic biology is an interdisciplinary science that involves using principles from disciplines such as engineering, molecular biology, cell biology, and systems biology. It involves remodeling existing organisms from nature or constructing completely new synthetic organisms for applications such as protein or enzyme production, bioremediation, value-added macromolecule production, and the addition of desirable traits to crops, to name a few.
Golden rice
Golden rice is a genetically modified...
Golden rice
Golden rice is a genetically modified...
4.7K
01:19MALDI-TOF Mass Spectrometry
4.6K
Mass spectrometry is a powerful characterization technique that can identify and separate a wide variety of compounds ranging from chemical to biological entities, based on their mass-to-charge ratio (m/z). The instruments that allow this detection, known as mass spectrometers, have three components: an ion source, a mass analyzer, and a detector. These spectrometers differ based on the nature of their ion source and analyzers.
Matrix-assisted laser desorption ionization (MALDI) is a commonly...
Matrix-assisted laser desorption ionization (MALDI) is a commonly...
4.6K